Automating Patient Triage in Radiology

Transcription

1 Improving Care and Efficiency through Analytics: Automating Patient Triage in Radiology Craig Froehle, PhD University of Cincinnati Lindner College of Business Dept. of Operations, Business Analytics & Information Systems College of Medicine Dept. of Emergency Medicine Cincinnati Children's Hospital Medical Center Anderson Center for Health Systems Excellence Collaborative work with Mark Halsted, MD and Neil Johnson, MD of CCHMC

2 The Setting Cincinnati Children s Hospital 587-bed private teaching pediatrics hospital Over 1.1 million patient encounters last year 16 patient care sites Consistently ranked in top 3 institutions Children s Radiology services Main hospital + 8 neighborhood locations Operate from a centralized stat box after hours Staffed by 1-2 radiologists (attendings, fellows, residents)

3 CCHMC locations with radiology services Main

4 Different imaging modalities X-ray MRI Ultrasound CT Cases Arrive Randomly Different requisition-delivery mechanisms Faxed from remote locations Brought by hand from on-site staff

5 Overall Goals: Ensure most critical patients are served first Reduce duration and variability of patient waiting Approach: Develop automated workflow management system Two functions: 1) Automatic triage of waiting cases 2) Automatic case routing and documentation of flow through the process

6 Measuring Baseline Performance End Procedure 1 Radiologist Dictates Radiologist Signs Off Findings Conveyed Overall 55 Emergency 23 Outpatient 57 Inpatient 103 Median (minutes) Baseline Sample: 6,093 exams, spanning 14 days Key performance metric How can we improve this?

7 Automating Triage

8 Automating Triage Radiologists use internal heuristics to select their next case Can we develop an algorithm to emulate their decision-making? Using easily obtainable data Simple to program

9 9 Potentially Influential Variables Patient Age Exam Type 20 exam categories Subjective Acuity Extreme, Mod., Mild Medical Acuity 5 categories (Airway, Trauma, Fracture, Pneum., Routine) Patient Anxiety High, Low Referring MD Anxiety High, Low Additional View? Yes, No Patient Waiting? Yes, No History Brief background

10 Data Collection Constructed 25 sets of 20 hypothetical cases Randomly generated Validated for OK medicine For each case, asked radiologists to rate (1-100) the urgency of the case Then asked to rank the 5 most urgent cases 22 radiologists (88%) participated

11 Patient/Case Information Please Provide the Following: Case # Patient Age Type Subjective Acuity Medical Acuity Patient Waiting Patient Anxiety Ref g MD Anxiety Add'l View? History 1 18 wk Chest Mild Pneum No Low High No Shortness of breath for 2 days 2 4 mo Chest Extreme Trauma Yes High High No MVA 1 hour ago Urgency Score (100 = Extreme 1 = None) Rank 5 Most Urgent 3 9 yr Abd Moderate Routine No High High No Abdominal pain 4 18 mo Chest Mild Airway No Low Low Yes cough 5 6 yr Knee Extreme Fracture Yes Low High No Fall on playground 4 hours ago 6 17 yr Chest Extreme Trauma Yes High High Yes MVA 7 5 yr Abd Extreme Routine Yes Low Low No Acute onset abdominal pain 8 9 yr Rad/Ulna Extreme Fracture No Low High No Arm bent after soccer collision 9 5 wk Femur Extreme Fracture No High High No Fell off changing table yr Knee Moderate Routine Yes High High No Knee pain yr Tib/Fib Mild Routine No Low High No Lump adjacent to tibia yr Foot Moderate Routine Yes Low Low No Stepped on nail 3 days ago, still has pain yr L Spine Extreme Trauma Yes High Low Yes Fell off horse back pain mo Chest Mild Pneum No Low High No cough yr Skull Mild Trauma Yes Low High Yes Bike accident 16 7 yr Chest Mild Trauma Yes Low High No Near drowning 17 6 yr Femur Mild Trauma Yes High High No Fall from tree mo Airway Extreme Airway Yes Low Low Yes Severe stridor mo Chest Mild Airway No Low Low Yes cough yr Ankle Moderate Trauma Yes Low Low No Soccer collision

12 Patient/Case Information Please Provide the Following: Case # Patient Age Type Subjective Acuity Medical Acuity Patient Waiting Patient Anxiety Ref g MD Anxiety Add'l View? History 1 18 wk Chest Mild Pneum No Low High No Shortness of breath for 2 days 2 4 mo Chest Extreme Trauma Yes High High No MVA 1 hour ago Urgency Score (100 = Extreme 1 = None) Rank 5 Most Urgent 3 9 yr Abd Moderate Routine No High High No Abdominal pain 4 18 mo Chest Mild Airway No Low Low Yes cough 5 6 yr Knee Extreme Fracture Yes Low High No Fall on playground 4 hours ago 6 17 yr Chest Extreme Trauma Yes High High Yes MVA 7 5 yr Abd Extreme Routine Yes Low Low No Acute onset abdominal pain 8 9 yr Rad/Ulna Extreme Fracture No Low High No Arm bent after soccer collision 9 5 wk Femur Extreme Fracture No High High No Fell off changing table yr Knee Moderate Routine Yes High High No Knee pain yr Tib/Fib Mild Routine No Low High No Lump adjacent to tibia yr Foot Moderate Routine Yes Low Low No Stepped on nail 3 days ago, still has pain yr L Spine Extreme Trauma Yes High Low Yes Fell off horse back pain mo Chest Mild Pneum No Low High No cough yr Skull Mild Trauma Yes Low High Yes Bike accident 16 7 yr Chest Mild Trauma Yes Low High No Near drowning 17 6 yr Femur Mild Trauma Yes High High No Fall from tree mo Airway Extreme Airway Yes Low Low Yes Severe stridor mo Chest Mild Airway No Low Low Yes cough yr Ankle Moderate Trauma Yes Low Low No Soccer collision

13 Patient/Case Information Please Provide the Following: Case # Patient Age Type Subjective Acuity Medical Acuity Patient Waiting Patient Anxiety Ref g MD Anxiety Add'l View? History 1 18 wk Chest Mild Pneum No Low High No Shortness of breath for 2 days 2 4 mo Chest Extreme Trauma Yes High High No MVA 1 hour ago Urgency Score (100 = Extreme 1 = None) Rank 5 Most Urgent 3 9 yr Abd Moderate Routine No High High No Abdominal pain 4 18 mo Chest Mild Airway No Low Low Yes cough 5 6 yr Knee Extreme Fracture Yes Low High No Fall on playground 4 hours ago 6 17 yr Chest Extreme Trauma Yes High High Yes MVA 7 5 yr Abd Extreme Routine Yes Low Low No Acute onset abdominal pain 8 9 yr Rad/Ulna Extreme Fracture No Low High No Arm bent after soccer collision 9 5 wk Femur Extreme Fracture No High High No Fell off changing table yr Knee Moderate Routine Yes High High No Knee pain yr Tib/Fib Mild Routine No Low High No Lump adjacent to tibia yr Foot Moderate Routine Yes Low Low No Stepped on nail 3 days ago, still has pain yr L Spine Extreme Trauma Yes High Low Yes Fell off horse back pain mo Chest Mild Pneum No Low High No cough yr Skull Mild Trauma Yes Low High Yes Bike accident 16 7 yr Chest Mild Trauma Yes Low High No Near drowning 17 6 yr Femur Mild Trauma Yes High High No Fall from tree mo Airway Extreme Airway Yes Low Low Yes Severe stridor mo Chest Mild Airway No Low Low Yes cough yr Ankle Moderate Trauma Yes Low Low No Soccer collision

14 Test #1: Intra-Physician Consistency 100 Corr overall = Case Case 4

15 Test #2: Inter-Physician Consistency 3 c 1 u pc u c Sum of Absolute Deviations from Group Mean Urgency Rating (Over 3 Cases) Mean Urgency Rating (Over 3 Cases)

16 Physician Selection Identified 5 representative docs: - Consistent decision-making - Within range of the majority - Highly experienced Case Case These 5 radiologists responses were then used for the algorithm development step

17 Variable Management Compared urgency means and distributions across categories; some were combined: Exam Type: 20 categories reduced to 2 Medical Acuity: 5 categories reduced to 2 Age: continuous variable dichotomized (<2, 2+) Average urgency rating add age graph here < Age (yr)

18 Constructing the Triage Algorithm Stepwise OLS regression using 5 radiologists responses: URGENCY = * SUBJACU (.36) * PATWAIT (.13) * REFANX (.08) * PATANX (.05) * DUMTYPE (.05) * DUMYOUNG (.01) F=35.52 (P<.0001) R 2 =.70 Not included: DUMMEDAC ADDVIEW But how well did it match our radiologists heuristics?

19 Testing the Triage Algorithm Prediction of rankings is primary metric: Corr = Act-Pre # % 18 Predicted ranking (algorithm) Actual ranking (radiologists)

20 Validation Survey

21 Validating the Triage Algorithm Provided each of the 5 radiologists with a set of 10 randomly generated, pre-ranked cases Found that: 3 of 5 docs made no changes or only swapped a single pair of adjacent cases (e.g., 3 rd 4 th ) 87% of all suggested changes were 1 or 2 places Only two large changes: -4 and +5 (same doc) Often used histories to substantiate changes We re still missing a key operational component

22 How to include patients waiting time? Physician and department beliefs: Stat patients: Should not wait >1 hour A short (~10 minutes) initial wait should not affect queue position Nonstat patients: Should generally be served after stat patients Can get lost among fast-moving stat cases

23 Incorporating Wait Times Started with the Standard Normal CDF Stat Cases t in minutes Time adder = Scale Factor 60.7( ) 250e t 2 Shape Factor Nonstat Cases t in hours Time adder = 60.1( ) 200e t 2

24 Wait Time Adder Acuity Score Time Adder 120

25 The Final Triage Algorithm URGENCY = * SUBJACU * PATWAIT * REFANX * PATANX * DUMTYPE * DUMYOUNG [3 levels] + Wait Time Adder {Stat or Nonstat} Urgency scores range for stat and for nonstat

26 Implementation

27 RadStream: Radiology Workflow Management

28 Radiologists Sign in to Services

29 Radiologists Select, Assign, & Review Cases Cases are pre-sorted per the triage algorithm Physicians may still select any case in their service

30 Completed Cases Automatically Routed to Call Center

31 Results

32 Changes to Workflow Tech answers 5 questions during imaging session Paper requisitions eliminated Waiting exams automatically triaged (sorted) Enhanced visibility and coordination Improved load-leveling across radiologists Expanded documentation of communications

33 Overall Goals: Ensure most critical patients are handled first Reduce duration and variability of patient waiting End Procedure 1 Radiologist Dictates Radiologist Signs Off Findings Conveyed Overall 55 Emergency 23 Outpatient 57 Inpatient 103 Median (minutes) Baseline Sample: 6,093 exams, spanning 14 days

34 Overall Goals: Ensure most critical patients are handled first Reduce duration and variability of patient waiting End Procedure 1 Radiologist Dictates Overall Emergency Outpatient Inpatient Median (minutes) Post-implementation Sample: 7,493 exams, spanning 15 days

35 Overall Goals: Ensure most critical patients are handled first Reduce duration and variability of patient waiting End Procedure 1 Radiologist Dictates Overall Emergency Outpatient Inpatient Median (minutes) Std. Dev. (minutes)

36 Physician Interruptions Decreased 15.1 Baseline Post-Implementation 11.8 Inter-Arrival (min.) Duration (min.) Significantly different at P<.05

37 Physician Interruptions Decreased 15.1 Baseline Post-Implementation 11.8 Inter-Arrival (min.) Duration (min.) Significantly different at P<.05

38

39

40 Conclusions for Care Delivery Decision-making in healthcare settings isn t always objective or rational Automating operational decision-making can be powerful But sometimes the data you need don t exist The benefits of efficiency are multiplicative

41

42 Improving Care and Efficiency through Analytics: Automating Patient Triage in Radiology Craig Froehle, PhD University of Cincinnati Lindner College of Business Dept. of Operations, Business Analytics & Information Systems College of Medicine Dept. of Emergency Medicine Cincinnati Children's Hospital Medical Center Anderson Center for Health Systems Excellence